What happened
Hugging Face’s GitHub release for LeRobot v0.7.0, published on November 5, describes a rebuild of installation, training, inference and the hardware layer. The release marks a breaking change from v0.6.1 to v0.7.0. On Linux, PyTorch and torchvision now come from PyPI as CUDA 13 builds and require NVIDIA driver 580.65 or newer; the cu128 index is removed. The notes direct users with older drivers to install PyTorch from a CUDA 12 index first. macOS 14 or later is required, with macOS 15 or later for MPS. The release also changes dependency pins, including Transformers 5.22 and Hugging Face Hub at least 1.31 but below 3.
Training changes affect tracking, distributed execution and fine-tuning. Trackio replaces Weights & Biases: the training extra installs Trackio, --wandb.* options become --trackio.*, and remote jobs using tracking need --trackio.space_id. The release says entity and mode have no equivalent, and checkpoints are no longer uploaded as artifacts by default. FSDP1 is replaced by FSDP2/HSDP, with new parallelism and accelerator options. FSDP checkpoints saved with LeRobot 0.6.x cannot be resumed, although their model.safetensors still loads. Accelerate YAML and environment configuration is rejected unless LEROBOT_ALLOW_ACCELERATE_ENV=1. Setting --policy.use_peft=true alone now trains LoRA rather than silently performing full fine-tuning. Several image transforms that had been no-ops on uint8 frames now apply, and ImageTransforms accepts tensors only.
Inference and rollout interfaces also change. The lerobot.async_inference module and lerobot[async] extra are removed; the release directs users to lerobot[remote], lerobot-policy-server and lerobot-rollout --inference.type=remote. The rollout inference module moves to lerobot.inference. Direct calls to select_action and predict_action_chunk may now return a PolicyOutput containing action and prediction rather than a bare tensor; the notes point to policy_output_to_transition for callers that need a transition. Simulation integrations move under lerobot.sim.integrations, environment configuration names change, and in-process environments return canonical observations with HWC uint8 images and flat float32 state.
The release lists new training and data capabilities alongside those breaking changes. Training additions include FSDP2/HSDP, gradient accumulation and DCP checkpoints; optional EMA of policy weights is enabled with --ema.enable=true. A new disk-less episode-pool reader is used by --dataset.streaming=true, with --dataset.streaming_* options. LeRobot adds native support for datasets stored as LanceDB datasets. The changed-items list also names policies and policy work including LaWAM, FLUX 3 Action, EO1 language runtime and text supervision, and a recipe-driven language-supervision feature. For hardware, Feetech and Dynamixel use a shared rustypot-backed SerialMotorsBus; their extras install rustypot rather than vendor SDKs, and earlier bus classes become deprecated aliases. The release lists no price for the software.
What the archive adds
New Runtime covered LeRobot documentation support for Dexmal’s DM0.5 VLA policy on October 23, 2026, in an earlier note. That note concerned support for a particular policy and its fine-tunable checkpoint; v0.7.0 is a new, broader release spanning the toolkit’s training, inference, data and hardware layers. This release is a new event on a known LeRobot thread, not simply a restatement of the earlier policy-support update.
Trend signal
This does not yet fit a pattern New Runtime tracks. The release is concrete evidence of a robotics toolkit’s expanding capabilities and changing interfaces, but it does not by itself establish one of the tracked patterns as a broader industry direction.
Why now
The release is dated November 5, 2026, and explicitly identifies changes from v0.6.1 that users need to address when upgrading. That makes timing operationally relevant for teams planning a version change: the CUDA driver requirement, training tracker migration and FSDP checkpoint limitation can affect whether an existing setup is ready to move.
Why it matters
For a team building robot-learning workflows, a unified toolkit release can make it easier to move between dataset handling, training, simulation and hardware. The new LanceDB support and disk-less streaming reader expand data options, while the changed training and remote-inference paths give teams new interfaces to adopt. But the breadth of this release also increases upgrade risk: CUDA requirements can block installation, tracker migration changes how runs are recorded, and older FSDP checkpoints cannot be resumed under the new path. A team should treat v0.7.0 as a coordinated workflow migration, not a routine dependency bump. Test the exact environment and checkpoint path first, then update training and rollout code against the new interfaces before moving robot operations over.
Limits and uncertainty
The release notes document interface changes and list features, but they do not provide comparative measurements showing how the new training, data or inference paths perform against the previous version. The release does not establish which combinations of policies, datasets, simulators and physical robots have been independently validated together. The listed CUDA and macOS requirements are concrete, but teams still need to verify their own drivers, hardware and dependencies. The release also describes new policy support without quantifying task success or generalization. Those questions matter when deciding whether a new capability is ready for production rather than experimentation.
Skill to build now
Build a small upgrade checklist around one representative LeRobot workflow: record the current driver and OS, training-tracker settings, checkpoint format, dataset backend, and rollout command. In a separate environment, test installation and resume behavior, then adapt tracking and remote inference to the new interfaces. This turns the release’s breaking changes into explicit pass/fail checks before a wider migration.
Reader takeaway
LeRobot 0.7.0 is a substantial robotics-toolkit release, but its most immediate consequence is migration work. New data and policy capabilities arrive alongside changed CUDA, training, checkpoint and inference paths. Inventory those dependencies and test a representative workflow before upgrading production runs.